lock free queue的实现

Arrray lock free queue

#ifndef __ARRAY_LOCK_FREE_QUEUE_IMPL_MULTIPLE_PRODUCER_H__
#define __ARRAY_LOCK_FREE_QUEUE_IMPL_MULTIPLE_PRODUCER_H__
#include "atomic_ops.h"
#include <assert.h> // assert()
#include <sched.h>  // sched_yield()
#include "logger_factory.h"
#include "mem_pool.h"
#include "shm_allocator.h"

/// @brief implementation of an array based lock free queue with support for 
///        multiple producers
/// This class is prevented from being instantiated directly (all members and
/// methods are private). To instantiate a multiple producers lock free queue 
/// you must use the ArrayLockFreeQueue fachade:
///   ArrayLockFreeQueue<int, 100, ArrayLockFreeQueue> q;


#define _WITH_LOCK_FREE_Q_KEEP_REAL_SIZE

template <typename ELEM_T, typename Allocator = SHM_Allocator>
class ArrayLockFreeQueue
{
   
    // ArrayLockFreeQueue will be using this' private members
    template <
        typename ELEM_T_, 
        typename Allocator_,
        template <typename T, typename AT> class Q_TYPE
        >
    friend class LockFreeQueue;

private:
    /// @brief constructor of the class
    ArrayLockFreeQueue(size_t qsize = LOCK_FREE_Q_DEFAULT_SIZE);
    
    virtual ~ArrayLockFreeQueue();
    
    inline uint32_t size();
    
    inline bool full();

    inline bool empty();
    
    bool push(const ELEM_T &a_data);   
    
    bool pop(ELEM_T &a_data);
    
    /// @brief calculate the index in the circular array that corresponds
    /// to a particular "count" value
    inline uint32_t countToIndex(uint32_t a_count);

    ELEM_T& operator[](unsigned i);
    
private:    
    size_t Q_SIZE;
    /// @brief array to keep the elements
    ELEM_T *m_theQueue;

    /// @brief where a new element will be inserted
    uint32_t m_writeIndex;

    /// @brief where the next element where be extracted from
    uint32_t m_readIndex;
    
    /// @brief maximum read index for multiple producer queues
    /// If it's not the same as m_writeIndex it means
    /// there are writes pending to be "committed" to the queue, that means,
    /// the place for the data was reserved (the index in the array) but  
    /// data is still not in the queue, so the thread trying to read will have 
    /// to wait for those other threads to save the data into the queue
    ///
    /// note this is only used for multiple producers
    uint32_t m_maximumReadIndex;

#ifdef _WITH_LOCK_FREE_Q_KEEP_REAL_SIZE
    /// @brief number of elements in the queue
    uint32_t m_count;
#endif
   
    
private:
    /// @brief disable copy constructor declaring it private
    ArrayLockFreeQueue<ELEM_T, Allocator>(const ArrayLockFreeQueue<ELEM_T> &a_src);

};


template <typename ELEM_T, typename Allocator>
ArrayLockFreeQueue<ELEM_T, Allocator>::ArrayLockFreeQueue(size_t qsize):
    Q_SIZE(qsize),
    m_writeIndex(0),      // initialisation is not atomic
    m_readIndex(0),       //
    m_maximumReadIndex(0) //
#ifdef _WITH_LOCK_FREE_Q_KEEP_REAL_SIZE
    ,m_count(0)           //
#endif
{
   
    m_theQueue = (ELEM_T*)Allocator::allocate(Q_SIZE * sizeof(ELEM_T));

}

template <typename ELEM_T, typename Allocator>
ArrayLockFreeQueue<ELEM_T, Allocator>::~ArrayLockFreeQueue()
{
   
    // std::cout << "destroy ArrayLockFreeQueue\n";
    Allocator::deallocate(m_theQueue);
    
}

template <typename ELEM_T, typename Allocator>
inline 
uint32_t ArrayLockFreeQueue<ELEM_T, Allocator>::countToIndex(uint32_t a_count)
{
   
    // if Q_SIZE is a power of 2 this statement could be also written as 
    // return (a_count & (Q_SIZE - 1));
    return (a_count % Q_SIZE);
}

template <typename ELEM_T, typename Allocator>
inline 
uint32_t ArrayLockFreeQueue<ELEM_T, Allocator>::size()
{
   
#ifdef _WITH_LOCK_FREE_Q_KEEP_REAL_SIZE

    return m_count;
#else

    uint32_t currentWriteIndex = m_maximumReadIndex;
    uint32_t currentReadIndex  = m_readIndex;

    // let's think of a scenario where this function returns bogus data
    // 1. when the statement 'currentWriteIndex = m_maximumReadIndex' is run
    // m_maximumReadIndex is 3 and m_readIndex is 2. Real size is 1
    // 2. afterwards this thread is preemted. While this thread is inactive 2 
    // elements are inserted and removed from the queue, so m_maximumReadIndex 
    // is 5 and m_readIndex 4. Real size is still 1
    // 3. Now the current thread comes back from preemption and reads m_readIndex.
    // currentReadIndex is 4
    // 4. currentReadIndex is bigger than currentWriteIndex, so
    // m_totalSize + currentWriteIndex - currentReadIndex is returned, that is,
    // it returns that the queue is almost full, when it is almost empty
    //
    if (countToIndex(currentWriteIndex) >= countToIndex(currentReadIndex))
    {
   
        return (currentWriteIndex - currentReadIndex);
    }
    else
    {
   
        return (Q_SIZE + currentWriteIndex - currentReadIndex);
    }
#endif // _WITH_LOCK_FREE_Q_KEEP_REAL_SIZE
}

template <typename ELEM_T, typename Allocator>